On Tue, Jan 8, 2019 at 12:26 AM Paolo Bonzini [off-list ref] wrote:

On 04/01/19 09:54, lantianyu1986@gmail.com wrote:

quoted

              rmap_head = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
                                        PT_PAGE_TABLE_LEVEL, slot);
-             __rmap_write_protect(kvm, rmap_head, false);
+             flush |= __rmap_write_protect(kvm, rmap_head, false);

              /* clear the first set bit */
              mask &= mask - 1;
      }
+
+     if (flush && kvm_available_flush_tlb_with_range()) {
+             kvm_flush_remote_tlbs_with_address(kvm,
+                             slot->base_gfn + gfn_offset,
+                             hweight_long(mask));

Mask is zero here, so this probably won't work.

In addition, I suspect calling the hypercall once for every 64 pages is
not very efficient.  Passing a flush list into
kvm_mmu_write_protect_pt_masked, and flushing in
kvm_arch_mmu_enable_log_dirty_pt_masked, isn't efficient either because
kvm_arch_mmu_enable_log_dirty_pt_masked is also called once per word.

Yes, this is not efficient.

I don't have any good ideas, except for moving the whole
kvm_clear_dirty_log_protect loop into architecture-specific code (which
is not the direction we want---architectures should share more code, not
less).

kvm_vm_ioctl_clear_dirty_log/get_dirty_log()  is to get/clear dirty log with
memslot as unit. We may just flush tlbs of the affected memslot instead of
entire page table's when range flush is available.

Paolo

quoted

+             flush = false;
+     }
+

--
Best regards
Tianyu Lan